Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Causal Covariate Shift Correction using Fisher information penalty

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes C3, a training-time penalty built from Fisher Information, and reports that it corrects covariate shift caused by batch order, improving accuracy by up to 20.3% over cross-validation.

desk verdict The C3 paper reports consistent accuracy gains from a Fisher-information penalty, but the claimed covariate-shift mechanism is not in the equations: the penalty is a curvature regularizer, not a density-shift correction. read the letter →

arxiv 2502.15756 v1 pith:TYKWA2EC submitted 2025-02-11 cs.LG

Behraj Khan , Behroz Mirza , Tahir Syed This is my paper · ORCID
classification cs.LG
keywords covariateshiftFisherinformationdistributioncontinuallearningregularizationcross-validationimportanceweightingdeep
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Standard training assumes an i.i.d. dataset, but continual and streaming learners receive batches whose feature densities drift in a fixed causal order, which biases cross-validation. The paper tries to establish that this 'causal covariate shift' can be corrected by a single penalty: the Fisher Information of the model's current parameters, added to the loss of every subsequent batch so that knowledge of earlier batches' densities constrains later updates. The authors report that this loss, called C3, improves accuracy over cross-validation by up to 20.3% in batchwise experiments and improves over the dynamic importance weighting state of the art by up to 12.9% when the full dataset is available, across 13 image benchmarks and 27 tabular benchmarks. If true, the result matters because the penalty uses derivatives already computed in backpropagation, so continual and federated learners get a shift correction at no change in complexity class.

What carries the argument

The load-bearing object is the Fisher Information Matrix, $I(\theta) = \mathbb{E}\left[-\frac{\partial^2 \log p(X\mid\theta)}{\partial\theta\,\partial\theta^\top}\right]$, used both as an approximation to the estimator covariance in a Gaussian KL-divergence estimate (via the Cramér-Rao bound $V(\hat\theta) \succeq I^{-1}(\theta)$) and as the penalty term itself. The mechanism is a weighted addition of this second-derivative term to the cross-entropy loss, with strength $\lambda$, so later batches are pulled toward the curvature geometry of earlier batches. The paper claims the penalty does not alter the algorithm's complexity class because it involves derivative operations already present in gradient descent, and it calibrates $\lambda$ experimentally, reporting results at $\lambda = 0.1$.

What would settle it

Take the same datasets and the same C3 loss, but replace the Fisher Information penalty with the Hessian of the current batch's loss, or with a fixed-norm penalty of comparable scale. If the reported accuracy gains do not change, the specific information content of the Fisher Information is not what carries the result, and the density-shift explanation is falsified.

Watch

Extended reading notes

Core claim

The central discovery claimed is that the KL divergence between the distribution of parameters implied by an earlier batch and the distribution implied by the current model can be approximated by the Fisher Information Matrix, and that this quantity can be inserted directly as a Tikhonov-style penalty in the loss. The derivation approximates the posterior over parameters as Gaussian with covariance $V(\hat\theta)$, invokes the Cramér-Rao lower bound $V(\hat\theta) \succeq I^{-1}(\theta)$, and uses a second-order Taylor expansion whose leading term is $I(\theta)$. The resulting loss is the cross-entropy loss with a $\lambda$-weighted Fisher Information penalty; the paper argues that because Fisher information is bounded above by entropy, penalizing it keeps the model's density knowledge from drifting. The experiments on image and tabular benchmarks are offered as evidence that the penalty raises accuracy under batchwise and foldwise causal shift, and also improves accuracy on natural covariate-shift benchmarks.

Load-bearing premise

The method assumes that the Fisher Information of the current model parameters, a local curvature quantity, actually represents the density shift between the previous and current batches and carries the earlier batch's distribution; if this identification fails, the method is a generic curvature regularizer rather than a covariate-shift correction.

Editorial extensions

If this is right

  • If C3 works as claimed, batchwise training under covariate shift can skip explicit density-ratio estimation; a Fisher Information penalty computed from the current parameters replaces it.
  • The reported gains grow as the number of batches shrinks (up to 20.3% at two batches on CIFAR-10), so the penalty is most valuable when each batch is a large, statistically reliable chunk of data.
  • The same loss improves accuracy on natural-shift benchmarks such as Kuzushiji-MNIST, CIFAR-10-C, and CIFAR-100-C, making batch-order correction double as a robustness correction.
  • Because no new derivative types are required, the penalty can be dropped into existing gradient-descent training loops for continual, federated, or AutoML pipelines at negligible extra cost.
  • Against dynamic importance weighting, C3 reports gains of 12.9% on Permuted-MNIST, 7.3% on CIFAR-10, and 5.1% on CIFAR-100, which the authors present as beating the full-dataset state of the art.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the identification of $V(\hat\theta)$ with $I^{-1}(\theta)$ is the load-bearing assumption. If it fails, the loss is a generic curvature regularizer and the reported gains need no covariate-shift story.
  • Editorial inference: a natural test the paper leaves open is permuting or reversing the batch order while keeping the same data and loss; if gains persist, the causal part of the claim would be weakened.
  • Editorial inference: because $\lambda$ is fixed at 0.1 for all reported results, sensitivity to penalty strength across datasets is underexplored; per-dataset calibration could materially change the headline numbers.
  • Editorial inference: because the model's memory of earlier batches is implicit in the current parameters' curvature, a running average of Fisher Information across batches is a natural extension that would make the accumulated density knowledge explicit and testable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes C3, a penalty term added to the cross-entropy loss during batchwise training, defined as the Fisher Information Matrix (FIM) of the current model and batch, scaled by a hyperparameter λ. The authors claim that this penalty accumulates knowledge about the data density of earlier training batches and corrects causal covariate shift, reporting accuracy improvements over cross-validation and importance-weighting baselines on a subset of image and tabular benchmarks.

Significance. If the mechanism were valid, a simple FIM-based penalty for batchwise or continual training would be practically valuable, and the claimed gains over importance-weighting baselines would be noteworthy. However, the derivation in Appendix A.1 does not connect the FIM penalty to any divergence between batch densities, and Algorithm 1 never uses the DKL values it computes. The method reduces to a curvature regularizer, so the central mechanistic claim is unsupported by the paper's own equations. The empirical evaluation also reports only 8 of 40 claimed datasets, with several baselines at near-chance accuracy, further weakening the evidence. The reported accuracy improvements may be real, but they cannot be attributed to the claimed covariate-shift-correction mechanism.

major comments (4)
  1. [Section 2, Eq. (3) and Appendix A.1, Eqs. (8)-(9)] The penalty term in Eq. (3) is I(θ) = -E[∂² log p(X|θ)/∂θ∂θᵀ], the FIM of the current batch's likelihood. Appendix A.1 derives an approximation to DKL(p(θ)||q(θ̂)) in Eq. (9) by replacing V(θ̂) with I⁻¹(θ), not with I(θ). Moreover, Eq. (9) is a parameter-space KL divergence between a true and estimated parameter distribution, not a divergence between two batch densities. The penalty actually used in Eq. (3) contains no term from any previous batch, so the central claim that C3 'accumulates knowledge about the data density of a training batch' is not realized by the equations presented.
  2. [Algorithm 1, lines 7-12] The algorithm computes DKL(Di,Dj) for every future pair (i,j) in line 7, but the loss L in lines 9-11 depends only on the cross-entropy term and the FIM penalty; the computed DKL values never enter L. Deleting lines 7-8 would change nothing, and the algorithm is invariant to reordering the batches apart from the ordinary sequential dependence of θ. Consequently, the method is at best a curvature-based regularizer, not a covariate shift correction, and the claimed mechanism is absent from the algorithm.
  3. [Appendix A.2, Eqs. (12)-(18)] This derivation contains mathematical errors and does not establish the claimed relationship between FIM and DKL. Eq. (12) defines DKL(P||Q) = ∫ P log(Q/P) dx, which is the negative of the standard KL divergence (a sign error). Eq. (16) leads to the conclusion I(θ) ≥ 0 in Eq. (18), which is trivially true by the definition of the FIM and carries no information about covariate shift. Thus the appendix does not provide theoretical support for using I(θ) as a penalty for distribution shift.
  4. [Experiments, Tables 1-3] Only 8 of the 40 claimed datasets are reported in the tables; there are no results for the other 32 KEEL datasets. The KMM baseline accuracies are near chance (e.g., 11.8% on MNIST in Table 1), suggesting a flawed implementation or improper tuning and undermining the SOTA comparison. The abstract's '12.9% over the full-dataset baseline' corresponds to ∆2 for Permuted-MNIST in Table 1, but the baseline definitions are inconsistent between tables, and no standard errors or significance tests are reported for the headline accuracy claims.
minor comments (6)
  1. [Section 1] There are typos such as 'covarite shift' and 'mostnmonly ascovarite shift' that should be corrected.
  2. [Section 2 and Appendix A] The notation DKL(P||Q) is used inconsistently: in Eq. (1) it is a parameter-space integral, in Eq. (7) it is a closed-form Gaussian KL, and in Appendix A.2 Eq. (12) it is an integral over X with a sign error. Please standardize the notation and fix the errors.
  3. [Algorithm 1] The pseudocode structure is ambiguous: the 'end for' pairing is unclear for lines 7-12, and the loss update in line 14 appears to be outside the inner loops. A clearer pseudocode with explicit loop termination would aid reproducibility.
  4. [Section B.6] The calibration of λ is presented only as a single figure (Fig. 1) with no numerical values, and the choice of λ = 0.1 is justified only by a claim of robustness. Please provide the actual calibration numbers or a table.
  5. [Section B.5] The list of 13 image datasets includes STL-10 and Tiny ImageNet, but no results for these datasets appear in any table; either report them or state why they are omitted.
  6. [Abstract and Section 4] The abstract claims '5.9% at minimum in foldwise benchmarks,' but no foldwise table with a 5.9% figure is identifiable in the paper; please point to the exact result or correct the claim.

Circularity Check

2 steps flagged · score 4.0 of 10

C3's accuracy gains are measured, but its covariate-shift mechanism is definitional: the FIM penalty contains no previous-batch term and the computed DKL is never used.

  1. self definitional [Section 2, Eq. 3 and Appendix A.1, Eq. 9]
    "Our method approximates the V (θ̂) by I(θ)... DKL(p(θ) ∥ q(θ̂)) ≈ ∫ p(θ) log[p(θ)/N(θ, I^{-1}(θ))] dθ ... L(x, y; θ) = −∫ P (y(x)) log(P (y|x; θ))dθ − λ × ∫ ∂² log p(X | θ)/∂θ∂θᵀ dθ"

    The derivation replaces V(θ̂) by I^{-1}(θ) in the covariance of q, but the penalty implemented in Eq. 3 is I(θ) = -E[∂² log p(X|θ)/∂θ∂θᵀ], evaluated on the current batch's likelihood X. This is a local curvature measure, not a divergence between two batch densities, and it contains no previous-batch term. Claiming that this penalty 'accumulates knowledge about the data density of a training batch' makes the covariate-shift mechanism true by definition of the penalty rather than by derivation.

  2. other [Appendix B, Algorithm 1, lines 7-11]
    "for j ← i + 1 to K do Compute DKL(Di, Dj) ... L(x, y; θ) = −∫ P (y(x)) log(P (y|x; θ)) dθ −λ × ∫ ∂² log p(X|θ)/∂θ∂θᵀ dθ"

    The DKL values computed for every future pair (Di, Dj) never appear in L or in the parameter update; L depends only on cross-entropy and the current-batch FIM. Deleting line 7 leaves the algorithm unchanged. Thus any claim that the loss penalizes deviations from previous batch distributions is by construction not present in the implemented objective.

full rationale

The empirical accuracy comparisons in Tables 1-3 are external measurements, not derived from the paper's equations, so the headline accuracy gains are not themselves circularly constructed. However, the paper's mechanistic claim that the FIM penalty 'accumulates knowledge about the data density of a training batch' is asserted by construction: Appendix A.1 replaces the covariance V(θ̂) in a KL-divergence approximation by I^{-1}(θ), and then Eq. 3 replaces the whole divergence term by I(θ) on the current batch's likelihood. No term from any earlier batch enters the loss, and Algorithm 1 computes DKL(Di, Dj) but never feeds it into L. The attribution of the empirical gains to covariate-shift correction therefore rests on a definitional identification of the FIM penalty with density-shift knowledge, not on a derivation. This is a partial circularity rather than a fully forced derivation, because the accuracy improvements are measured and independent of the derivation chain.

Assumptions & free parameters 1 free parameters · 5 assumptions · 1 invented entities

The paper's central claim rests on one tuned scalar lambda and on the unproven identification of the Fisher information of current parameters with the density shift between batches. The derivation in the appendix establishes only the standard local KL-FIM relation, not the validity of the penalty as a shift correction.

free parameters (1)
  • lambda (penalty strength) = 0.1
    Selected from {0.01, 0.04, 0.07, 0.1} based on Fig 1 because C3 was 'more robust' at 0.1; all reported results use this tuned value.
assumptions (5)
  • standard math Cramér-Rao lower bound V(theta-hat) >= I^{-1}(theta) justifies replacing the covariance in the Gaussian KL approximation by the inverse Fisher information.
    Used in Eq 5 and Eq 9 to substitute I^{-1}(theta) for V(theta-hat).
  • ad hoc to paper The Fisher information of the current model parameters is a valid proxy for the KL divergence between successive training batches and encodes the density of previous batches.
    This is the core mechanism asserted around Eq 3 and Algorithm 1; no derivation connects I(theta) to a between-batch density shift.
  • domain assumption The variational posterior is Gaussian: q(theta-hat) approximately N(theta, V(theta-hat)).
    Used in Eq 6 and Eq 8; no evidence is given that the posterior is Gaussian.
  • standard math Second-order Taylor expansion of the KL divergence, with higher-order terms dropped.
    Used in Appendix A.2, Eqs 15-16; valid for small parameter deviations but not justified for large batch shifts.
  • domain assumption 'Causal covariate shift' is defined as immutability of the batching sequence.
    Section 1 defines causality this way; this is a naming choice, not a causal model.
invented entities (1)
  • Causal covariate shift as defined by the immutability of the batching sequence
    purpose: Frames the batching order as a causal constraint so the method can claim to correct a new type of shift.
    The paper provides no formal causal model or falsifiable handle; it renames batched distribution shift as causal based on a citation to WaveNet.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Covariate Shift Correction using Fisher information penalty." pith.science (2026). https://pith.science/paper/TYKWA2EC

@misc{pith2026250215756,
  author       = {Pith},
  title        = {Pith review of: Causal Covariate Shift Correction using Fisher information penalty},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TYKWA2EC}},
  note         = {Machine review of arXiv:2502.15756}
}
abstract

Evolving feature densities across batches of training data bias cross-validation, making model selection and assessment unreliable (\cite{sugiyama2012machine}). This work takes a distributed density estimation angle to the training setting where data are temporally distributed. \textit{Causal Covariate Shift Correction ($C^{3}$)}, accumulates knowledge about the data density of a training batch using Fisher Information, and using it to penalize the loss in all subsequent batches. The penalty improves accuracy by $12.9\%$ over the full-dataset baseline, by $20.3\%$ accuracy at maximum in batchwise and $5.9\%$ at minimum in foldwise benchmarks.

Figures

Figures reproduced from arXiv: 2502.15756 by the authors.

Figure 1
Figure 1. Performance of C 3 for varying λ in foldwise (k = 5) settings. 7. Experimental design: To study the effect of causal covariate shift caused by fragmenta￾tion, we perform evaluations on datasets with natural covariate shift and also on clean (free of covariate shift) datasets. We use accuracy as the first and direct evaluation metric in all experiments. We run each experiment 5 times and report average results due to… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CalTwin: Towards Calibrated, Shift-Robust Medical World Models via Fisher-Information Regularisation

    cs.LG 2026-07 conditional novelty 4.0 of 10

    Combining FIM shift regularization with a confidence-misalignment penalty on a GRU latent dynamics model modestly cuts OOD next-step MSE on PhysioNet sepsis data but does not consistently transfer to eICU.

Reference graph

Works this paper leans on

30 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    Keel data-mining software tool: data set repository, integration of algorithms and experimental analysis framework

    Jes \'u s Alcal \'a -Fdez, Alberto Fern \'a ndez, Juli \'a n Luengo, Joaqu \' n Derrac, Salvador Garc \' a, Luciano S \'a nchez, and Francisco Herrera. Keel data-mining software tool: data set repository, integration of algorithms and experimental analysis framework. Journal of Multiple-Valued Logic & Soft Computing, 17, 2011

    work page 2011

  2. [2]

    Deep learning for classical japanese literature

    Tarin Clanuwat, Mikel Bober-Irizar, Asanobu Kitamoto, Alex Lamb, Kazuaki Yamamoto, and David Ha. Deep learning for classical japanese literature. arXiv preprint arXiv:1812.01718, 2018

    arXiv 2018

  3. [3]

    An analysis of single-layer networks in unsupervised feature learning

    Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp.\ 215--223. JMLR Workshop and Conference Proceedings, 2011

    2011

  4. [4]

    Sample selection bias correction theory

    Corinna Cortes, Mehryar Mohri, Michael Riley, and Afshin Rostamizadeh. Sample selection bias correction theory. In Algorithmic Learning Theory: 19th International Conference, ALT 2008, Budapest, Hungary, October 13-16, 2008. Proceedings 19, pp.\ 38--53. Springer, 2008

    work page 2008

  5. [5]

    Monotonicity of entropy and fisher information: a quick proof via maximal correlation

    Thomas A Courtade. Monotonicity of entropy and fisher information: a quick proof via maximal correlation. arXiv preprint arXiv:1610.04174, 2016

    arXiv 2016

  6. [6]

    Rethinking importance weighting for deep learning under distribution shift

    Tongtong Fang, Nan Lu, Gang Niu, and Masashi Sugiyama. Rethinking importance weighting for deep learning under distribution shift. Advances in neural information processing systems, 33: 0 11996--12007, 2020

    work page 2020

  7. [7]

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories

    Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Computer Vision and Pattern Recognition Workshop, 2004

    2004

  8. [8]

    An empirical investigation of catastrophic forgetting in gradient-based neural networks

    Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013

    arXiv 2013

Show all 30 references
  1. [9]

    Covariate shift by kernel mean matching

    Arthur Gretton, Alex Smola, Jiayuan Huang, Marcel Schmittfull, Karsten Borgwardt, Bernhard Sch \"o lkopf, et al. Covariate shift by kernel mean matching. Dataset shift in machine learning, 3 0 (4): 0 5, 2009

    2009

  2. [10]

    Benchmarking neural network robustness to common corruptions and perturbations

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019

    2019

  3. [11]

    Correcting sample selection bias by unlabeled data

    Jiayuan Huang, Arthur Gretton, Karsten Borgwardt, Bernhard Sch \"o lkopf, and Alex Smola. Correcting sample selection bias by unlabeled data. Advances in neural information processing systems, 19, 2006

    2006

  4. [12]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  5. [13]

    On information and sufficiency

    Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22 0 (1): 0 79--86, 1951

    1951

  6. [14]

    The mnist database of handwritten digits

    Yann LeCun. The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/, 1998

    1998

  7. [15]

    Study on the impact of partition-induced dataset shift on k -fold cross-validation

    Jose Garc \' a Moreno-Torres, Jos \'e A S \'a ez, and Francisco Herrera. Study on the impact of partition-induced dataset shift on k -fold cross-validation. IEEE Transactions on Neural Networks and Learning Systems, 23 0 (8): 0 1304--1312, 2012

    2012

  8. [16]

    Mnist-c: A robustness benchmark for computer vision

    Norman Mu and Justin Gilmer. Mnist-c: A robustness benchmark for computer vision. arXiv preprint arXiv:1906.02337, 2019

    1906 arXiv

  9. [17]

    Reading digits in natural images with unsupervised feature learning

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011

    2011

  10. [18]

    A new lower bound for kullback-leibler divergence based on hammersley-chapman-robbins bound

    Tomohiro Nishiyama. A new lower bound for kullback-leibler divergence based on hammersley-chapman-robbins bound. arXiv preprint arXiv:1907.00288, 2019

    1907 arXiv

  11. [19]

    Revisiting natural gradient for deep networks

    Razvan Pascanu and Yoshua Bengio. Revisiting natural gradient for deep networks. arXiv preprint arXiv:1301.3584, 2013

    2013 arXiv

  12. [20]

    Dataset shift in machine learning

    Joaquin Qui \ n onero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2008

    2008

  13. [21]

    Machine learning in non-stationary environments: Introduction to covariate shift adaptation

    Masashi Sugiyama and Motoaki Kawanabe. Machine learning in non-stationary environments: Introduction to covariate shift adaptation. MIT press, 2012

    2012

  14. [22]

    Covariate shift adaptation by importance weighted cross validation

    Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert M \"u ller. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8 0 (5), 2007 a

    2007

  15. [23]

    Direct importance estimation with model selection and its application to covariate shift adaptation

    Masashi Sugiyama, Shinichi Nakajima, Hisashi Kashima, Paul Buenau, and Motoaki Kawanabe. Direct importance estimation with model selection and its application to covariate shift adaptation. Advances in neural information processing systems, 20, 2007 b

    2007

  16. [24]

    Senior, and Koray Kavukcuoglu

    A \" a ron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew W. Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. CoRR, abs/1609.03499, 2016. URL http://arxiv.org/abs/1609.03499

    2016 arXiv

  17. [25]

    Adaptive filtering and hypothesis testing: Application to cancerous cells detection

    Vincent Vigneron, Tahir Q Syed, Georgia Barlovatz-Meimon, Michel Malo, Christophe Montagne, and Sylvie Lelandais. Adaptive filtering and hypothesis testing: Application to cancerous cells detection. Pattern recognition letters, 31 0 (14): 0 2214--2224, 2010

    2010

  18. [26]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms

    Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017

    2017 arXiv

  19. [27]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  20. [28]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  21. [29]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  22. [30]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.